The default answer to "where do we run our AI workloads" is a US hyperscaler — a quiet decision to hand the most strategic layer of your future business systems to a vendor whose roadmap, pricing, and jurisdiction you don't control. This talk presents a working blueprint for an alternative: an agent-serving platform built for AI resilience, where the choice of model, provider, and location is a configuration value rather than an architectural commitment.

At the heart of the platform is an AI gateway that unifies inference across every model — frontier, open-weight, or self-hosted — behind one interface, with routing, model swap by config, cost tracking, observability, and a built-in chat front-end. Around it sit the vector store, state and caching, a fine-tuning loop, and the full monitoring stack — all defined as code and provisioned automatically across cloud providers. The result: swapping a model or moving inference between providers becomes a configuration change rather than a rebuild — the dial that makes the platform sovereign, resilient, and trusted by design.

I'll show how it runs deployed across two independent cloud providers in a resilient configuration, where no single provider, model vendor, or jurisdiction is a point of failure. The gateway also doubles as an evaluation harness — letting you audition open models against a frontier baseline through the real workload before committing any infrastructure, so trust in a model is measured rather than assumed. I'll touch on the practical realities of running open models on your own GPUs, but keep the focus on the architecture that makes the choice reversible rather than the hardware underneath it.

The argument is not anti-hyperscaler. It's that the binary framing — managed or self-hosted, pick one — is the wrong way to decide. The platform that lets you change your mind cheaply, across providers and jurisdictions, is the one worth building. Especially for European organizations with regulated data and a real interest in not exporting it, that combination of sovereignty, resilience, and trust is worth more than the marginal convenience of a managed endpoint.